Partner For Good

Background

As of the 5th of December 2024, to contribute to the UN’s Sustainability Development Goals, the NI Government reviewed and revised The Procurement Act. This government policy aims to achieve social and environmental progress through government spending. Consequently, for service contracts exceeding £500,000, social value impact is considered one of the top 3 deciding factors for bids, alongside quality and cost (Department of Finance NI, 2024).

It is therefore crucial that a supplier — a business who supply their services to the government, hoping to bid for a government contract — demonstrate a strong element of social value in their bid. This can be achieved by partnering with brokers: third sector organisations with the skills and experience to support businesses in delivering their social value requirements (Social Value NI, 2023). My aim is to build a website which helps suppliers create more impactful social value contributions by facilitating connections with relevant and appropriate brokers.

Stakeholder feedback indicated that some brokers found the user experience of current solutions unsatisfactory, and that as small businesses, they were infrequently selected as a partner on the current platform. My AI search aims to help smaller brokers who reported they have never been contacted through the current platform to be found — because the search is based on relevance rather than big names or fancy branding.

Next.js

Next.js is an open-source framework built on top of React, popular for the high-performance sites it can create and fast build times powered by Rust-based JavaScript tooling (Vercel, 2024). Its file-based routing invalidates the need for react-router, and built-in tooling like Tailwind makes it easy to get started. The trade-off is a steep learning curve around its data-fetching paradigms and React Server Components.

I chose Next.js because I needed a framework that would scale well with fast load times. Server-side rendering, static site generation and prefetching links let me ensure the site loaded data and pages quickly — important on the directory pages, where I could be loading large amounts of data at once.

SST

SST v3 is an infrastructure-as-code framework, completely rebuilt since v2 to leverage Pulumi and Terraform instead of CDK and CloudFormation (V, 2024). Instead of generating CloudFormation templates, SST now uses AWS SDK calls to create resources, which makes it much faster. I used SST to define all of my AWS resources and create different deployment stages.

One advantage is that SST v3 is now cloud-agnostic: if AWS pricing became too expensive, I could move the infrastructure to Azure or GCP — a CDK approach would lock me to AWS. The reason I used SST specifically was OpenNext, an open-source tool that let me take advantage of Next.js features like server-side rendering while deploying on AWS services instead of Vercel (OpenNext, 2025). That gave me the flexibility of AWS, the novel features of Next.js, and a single dashboard for all my cloud resources.

AI Search with pgvector

pgvector is an open-source Postgres plugin that allows users to store vector data alongside traditional relational data, giving the ability to perform similarity searches directly in the table (pgvector, 2024). The biggest advantage is storing vector data as just another column rather than provisioning a separate vector store — reducing both cost and the number of API calls needed for a similarity search.

pgvector was an easy choice here. I needed hybrid search — only running a similarity search on brokers whose status is “active” — and with pgvector I could filter on other columns and return the full row in a single SQL command and API call. To enable the extension on my deployed Postgres database, I created a tunnel into the VPC and ran the SQL directly. I set up a 1024-dimension vector column for high accuracy without excessive storage, then a function using a Hierarchical Navigable Small World (HNSW) algorithm for Approximate Nearest Neighbour search (Malkov and Yashunin, 2016).

To search, I take a vector embedding and run a cosine similarity search, returning results above 10% similarity, ordered highest to lowest, top 3. I chose cosine similarity specifically because it works best for natural language — measuring similarity without consideration of the length of text (magnitude). A user could enter a very short prompt that is very similar to a broker; magnitude-sensitive measures like Squared Euclidean would not consider these similar (Cardenas, 2023).

Security

I use GitHub’s Dependabot to run Static Code Analysis scans weekly and open a PR to bump versions when a vulnerability is present — guarding against vulnerable and outdated components, A06 in the OWASP Top 10 (OWASP, 2021b). To prevent sensitive data leaks and A02 Cryptographic Failures (OWASP, 2021a), I use AWS Secrets Manager and GitHub Action Secrets so secrets are never stored as plaintext in code or in the compiled application. If anyone searched the compiled app — or my GitHub were ever compromised — they couldn’t use what they found to reach user data or my cloud services.

Accessibility

The application was built with a strong focus on accessibility, aiming to meet WCAG standards. I integrated accessibility checks into development with eslint-plugin-jsx-a11y, built into Next.js (Next.js, 2025a), which flags issues like missing alt text, incorrect aria attributes, or inputs without labels as I code. I used semantic HTML (nav, main, header, footer, button) for proper structure that matters to screen readers, made sure every form control has an associated label, and implemented full keyboard navigation with a logical tab order and clear focus indicators — accomplishing Principle 2.1 of the WCAG standards (W3C, 2023).